Research map / Architectures and efficiency / RNNs and normalization
CALANet: Cheap All-Layer Aggregation for Human Activity Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CALANet, a 1D CNN for wearable human activity recognition that aggregates features from all layers via learnable channel-wise transformation matrices and a scalable layer aggregation pool, keeping FLOPs comparable to shallow CNNs.
Problem
Real-time sensor-based human activity recognition (HAR) with lightweight CNNs suffers limited accuracy because the classifier only uses last-layer features; aggregating all layers normally raises computational cost.
Approach
Learnable channel-wise transformation matrices map each layer's features to a fixed-size vector with cost independent of temporal resolution, and grouped convolution with channel shuffle plus a fixed-size softmax input (SLAP) keep the total cost equivalent to shallow CNNs while allowing all-layer aggregation.
Key results
- CALANet achieved F1 of 97.5 on KU-HAR, 96.1 on UCI-HAR, and 98.2 on REALDISP, with the highest F1 among compared models on most datasets.
- CALANet reached 81.6 F1 on OPPORTUNITY and 79.4 on PAMAP2 at 19.3M and 74.9M FLOPs respectively, versus 82.2 for IF-ConvTransformer at 986.2M FLOPs on OPPORTUNITY.
- Average real-time inference time of CALANet was 2.25 ms per window on a Ryzen 7 5800X CPU, similar to Shallow ConvNet (2.15 ms) despite being nine layers deep.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CALANet instead of the strongest listed baseline | HAR models · UCI-HAR | F1: 95.8 -> 96.1 | Helped |
| Use CALANet instead of the strongest listed baseline | HAR models · UniMiB-SHAR | F1: 81.4 -> 78.3 | Hurt |
| Use CALANet instead of the strongest listed baseline | HAR models · DSADS | F1: 87.6 -> 90.0 | Helped |
| Use CALANet instead of the strongest listed baseline | HAR models · OPPORTUNITY | F1: 82.3 -> 81.6 | Hurt |
| Use CALANet instead of the strongest listed baseline | HAR models · KU-HAR | F1: 97.8 -> 97.5 | Hurt |
| Use CALANet instead of the strongest listed baseline | HAR models · PAMAP2 | F1: 80.1 -> 79.4 | Hurt |
| Use CALANet instead of the strongest listed baseline | HAR models · REALDISP | F1: 98.5 -> 98.2 | Hurt |
| Add SLAP (ablation) | CALANet · KU-HAR | F1: 93.8 -> 97.5 | Helped |
| Add SLAP (ablation) | CALANet · PAMAP2 | F1: 73.1 -> 79.4 | Helped |
| Use LCTMs instead of fully-connected layers (ablation) | CALANet with ALA · KU-HAR | F1: 95.0 -> 97.5 | Helped |
| Use LCTMs instead of fully-connected layers (ablation) | CALANet with ALA · PAMAP2 | F1: 72.8 -> 79.4 | Helped |
| Vary layer aggregation pool depth L (ablation) | CALANet · Seven HAR datasets | F1: UCI-HAR: 93.2, 93.9, 94.8, 96.1, 95.7; UniMiB-SHAR: 72.8, 76.0, 78.0, 78.3, 77.5; DSADS: 87.2, 84.5, 86.0, 90.0, 89.4; O | Mixed |
Surprising findings
- Adding CALA to SqueezeNet improved F1 on 71% of datasets while keeping FLOPs roughly unchanged.
- CALANet did not beat the best accuracy-oriented models on every dataset; on OPPORTUNITY, Shallow ConvNet and T-ResNet/MILLET were comparable or higher in some cases, and CALANet's MIT-BIH ECG accuracy (98.2) was below Pham et al. (98.5).
Limitations
- Does not account for varying computational budgets across devices or runtime/memory-access costs.
- Evaluated only on HAR datasets and one ECG dataset; ECG comparison is limited to a single dataset.
Experimental setup
Datasets: UCI-HAR UniMiB-SHAR DSADS OPPORTUNITY KU-HAR PAMAP2 REALDISP MIT-BIH arrhythmia
Benchmarks: UCI-HAR UniMiB-SHAR DSADS OPPORTUNITY KU-HAR PAMAP2 REALDISP MIT-BIH arrhythmia
Metrics: F1-score FLOPs inference time average accuracy
Baselines: Shallow ConvNet RepHAR Res-GTSNet DeepConvLSTM Bi-GRU-I RevAttNet IF-ConvTransformer T-ResNet T-FCN MILLET DSN
Builds on: ShuffleNet Res-GTSNet SqueezeNet
Closest papers by meaning
- Generalizable Sensor-Based Activity Recognition via Categorical Concept Invariant Learning (AAAI 2025 · similarity 0.82)
- Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition (AAAI 2025 · similarity 0.82)
- HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.82)
- UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.81)
- MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.80)
- ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.